A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 13:30 UTC
Change warning logs to debug for unimplemented MXFP4 Linear/Attention (#29441)
Signed-off-by: Michael Goin <mgoin64@gmail.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Michael Goin committed
Nov 25, 2025 at 17:52 UTC
7df0289782ab500b2713b7521979c28de2b21cac
1 file changed
+6
-4
vllm/model_executor/layers/quantization/mxfp4.py
+6
-4
@@ -196,9 +196,10 @@ class Mxfp4Config(QuantizationConfig):
196
# TODO: Add support for MXFP4 Linear Method.
197
# MXFP4 LinearMethod is available in AMD-Quark, refer to that implementation
198
# if you are interested in enabling MXFP4 here.
199
- logger.warning_once(
199
+ logger.debug_once(
200
"MXFP4 linear layer is not implemented - falling back to "
201
- "UnquantizedLinearMethod."
201
+ "UnquantizedLinearMethod.",
202
+ scope="local",
203
)
204
return UnquantizedLinearMethod()
205
elif isinstance(layer, FusedMoE):
@@ -208,9 +209,10 @@ class Mxfp4Config(QuantizationConfig):
209
return Mxfp4MoEMethod(layer.moe_config)
210
elif isinstance(layer, Attention):
211
# TODO: Add support for MXFP4 Attention.
211
- logger.warning_once(
212
+ logger.debug_once(
213
"MXFP4 attention layer is not implemented. "
213
- "Skipping quantization for this layer."
214
+ "Skipping quantization for this layer.",
215
+ scope="local",
216
)
217
return None
218